Generic output mixing in a dirty connector: erasure or saving?

[RossRichardson]

DONE / HANDOFF — claim RELEASED. PR2050, exact head a86b65e648f5b004f4aa68b3ed2c7975e8ca0163, is open for review. Source base c676548106012571fefa1dce0ec16f1bcfe11dce; integrated main e112d71c8d7bb32435debcc54df932593dd8223a, with all scoped inputs unchanged.

Established: for a fixed word uniform over independently arbitrary exact F2 readers, any scalar GL mixing on outputs plus any finite dirty output-only helper bank can be erased. The new word preserves the complete B timeline and does not increase the count of any A_i, even granting arbitrary physical reader ports in the mixed word. Thus output mixing cannot improve any fixed nonnegative port-independent complete map-call tariff. The inherited five-source/all 31-mask optimum remains 74 and the heterogeneous fixture remains 182, even with free mixing. This is an all-parameter written proof, not a formal-kernel proof or new kappa.

Why the hypotheses matter: a shared-reader broadcast breaks isolation and is deliberately rejected by the generic checker. A valid identity with three A calls shows that the old parity argument cannot apply directly to mixed words; the lower bound must transfer through erasure. Source/connector mixing, non-scalar block frames, intermediate observers and port-dependent total routing costs remain outside scope. No selected-supplier occurrence, complete physical compiler, cold replay or exponent transfer is supplied.

Evidence: registered CPython 3.12.12 replay passes. Separate normalized-coefficient and noncommutative/suffix-routing implementations agree on every dirty column and both orientations. Two 945-pairing censuses agree; all 168 GL(3,2) frames in the stated fixture and all 19,531 words of length<=6 in the stated small alphabet are checked, alongside 256 matrix pairs and 4,096 common-reader triples. Seven behavior-test groups pass, including actual detector removal, helper-only defects, missing/corrupt evidence, source drift, normal/optimized unchanged relocation and no-overwrite generation. Hosted CI is queued at publication; no hosted pass or human acceptance is claimed.

Reproduce: python3.12 -I -B scripts/research_package.py --package research/generic-output-mixing-erasure.

Next useful question: identify one pinned selected call family with an actual reader relation that defeats independent isolation, then price its two-output sharing identity including target access, routing and restoration. No same-model generic output-mixing search remains justified by this result.

Credit Ross Richardson PR943/934 and the retained dirty-commutator/fanout lineage, Adam Gall PR1915, Arham Garg PR1945, Cinna Davis PR1770, and 1337isnot1337 PR1707. Substantial OpenAI Codex assistance with separate recursive-interface and exact-verification specialists; Apache-2.0. Selector nonce 779e702f7c4ee535de2ee386c8e4f4bf, refined connector-alphabet-change scope. Final overlap 328 open PRs/691 Discussion records, complete pagination without warnings at 18:08–18:10 UTC, not a novelty guarantee. No merge, closure or selected-frontier change.